Gemini 4 Argon Test: 23 Tasks vs Opus 5.5 and GPT-6 Astra

Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.

👁 126 reads

On the evening of October 13, 2026, roughly half an hour after Google opened Gemini 4 Argon to paying customers, I ran it through the exact same battery of 23 tasks assigned a day earlier to Muse Wonder, Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra, and Gemini 3.8 Flash. Argon scored 99.5%, cost $0.22, and finished faster than any competitor: 135 seconds for the complete run. Its only deducted point stemmed from an evaluation script bug, although a genuine issue surfaced just a few paragraphs down in the same draft. You can find model specs, pricing, and availability details in the guide to Gemini 4 Argon.

Benchmark results across all models

The overall score represents the unweighted mean across all 23 tasks. For multi-part tasks, the percentage reflects the share of criteria met. Among flagship models, Argon was the cheapest: 3.4x cheaper than Claude Opus 5.5, 5x cheaper than GPT-6 Astra, and 2.9x cheaper than Muse Wonder. Only Gemini 3.8 Flash came in lower, sitting in a budget tier.

Tasks evaluated in the benchmark

Performance breakdown by category

Bold formatting highlights the top score per category, breaking ties with the lowest cost. The remaining models dropped marks elsewhere: Muse Wonder stumbled on code (98.4%, outputting a nonexistent Polish numeral construction), Gemini 3.8 Flash also dropped code points (96.9%, treating Christmas Eve as a holiday starting in 2024 instead of 2025), while Claude Fable 5.1 and Opus 5.5 lost points in Polish language tasks (96.4%, breaching the 60-word limit on the summary).

Speed and cost advantages

Argon produces concise responses without delay. At the low reasoning tier, OpenRouter reported zero reasoning tokens across the entire 23-task run, and the 10-error proofreading prompt returned in 2.6 seconds. In contrast, Muse Wonder spent 636 reasoning tokens on the travel task alone. Key observations:

• Lisbon itinerary: 4 tool calls across 2 turns, dodging both traps (an overly early flight and a poorly rated hotel), calculating roughly $448 for the budget, and scheduling a museum visit due to an 80% rain forecast. Matches the competition, but at a cost of only $0.011.
• Retail return request: Correctly targeted the right order item, rejected the return, and explained to the customer that 32 days had elapsed past the 30-day window.
• EUR invoices: Successfully retried after the simulated 503 error, excluded paid invoices, and calculated the correct total.
• Long document processing: 5 out of 5 correct answers in 4.5 seconds for $0.10. It noted that a warehouse footprint had been revised to 6,200 m2 and answered that no TikTok campaign existed. Only GPT-6 Astra answered faster (3.5 s), but cost $0.69.

Token efficiency is another major factor on non-English text. Those 158,111 Polish characters converted to 51,482 tokens for Argon versus 87,675 for Claude, a 70% overhead. Combined with cheaper base pricing, querying 80 meeting transcripts cost $0.10 on Argon, $0.19 on Muse Wonder (46,929 tokens at $4 per million), $0.35 on Opus 5.5, $0.69 on Astra, and $0.88 on Fable 5.1.

The legal appeal test: false penalty and real hallucination

The automated script initially awarded Argon 8 out of 9 points on a legal disability benefit appeal letter, flagging a missing recipient line for the regional court. Inspecting the raw text showed the recipient was present; the model had simply used standard Polish genitive case phrasing rather than the nominative string expected by the script. Correcting that manual false negative brings Argon to 100%, matching GPT-6 Astra. The automated 99.5% is kept in the comparison table to maintain uniform scoring rules across all runs.

The draft handled procedural specifics well, noting the prerequisite objection step and calculating the statutory deadline around weekend and holiday closures to land on November 2, 2026. However, an unprompted hallucination appeared in the factual narrative. Given only a basic prompt stating that an evaluator overlooked 2026 cardiology treatment records, Argon fabricated specific physical symptoms and misconduct accusations against the physician (translated from the tested Polish output):

“The symptoms of the condition, such as rapid fatigue, exertional dyspnea, blood pressure spikes, and chest pain, prevent me from functioning normally in everyday life (…)”

“The certifying physician, when issuing the decision, conducted the examination in a perfunctory manner (…)”

Opus 5.5 also suggested fatigue and shortness of breath, but wrapped them inside placeholder brackets for the user to verify. Argon inserted them directly as established facts. In formal filings subject to evidentiary review, fabricating medical claims is dangerous. Any administrative drafts created by Argon require line-by-line review.

Pricing changes after December 1

The launch pricing of $2 / $10 per million tokens runs through November 30, 2026. Starting December 1, Argon shifts to $4 / $20, bringing its listed rates level with Opus 5.5 and Muse Wonder. Applying those base rates roughly doubles the cost of this suite:

Even with list prices equalized, Argon remains roughly 1.7x cheaper than Opus 5.5 on this benchmark. That delta stems from direct output without excess hidden thinking tokens and a denser tokenizer on non-English text. Context caching rates also adjust from $0.10 to $0.20 per million tokens.

Test caveats and scope

• Effort=low. All models were tested on their lowest reasoning tier to measure cost-efficient throughput. Google’s official benchmarks use maximum reasoning effort, reflecting different operating conditions.
• Single test pass. 23 distinct tasks, executed once. Single-point discrepancies reflect margin of error rather than conclusive capability gaps.
• Test execution dates. Competing models ran on October 12; Argon ran on October 13 shortly after public API launch. Prompts and evaluation code were identical, but network conditions and server load varied.
• Performance ceiling. All leading models clustered between 99% and 100%. This suite evaluates latency, cost efficiency, and specific failure modes rather than frontier-level puzzle solving.
• Automated scoring. Automated evaluations focus on factual matching and structured checks. The appeal letter shows that scripts can misread inflected language, which is why all failed assertions were inspected manually.
• OpenRouter routing. Evaluated via OpenRouter rather than direct Google Cloud endpoints. Model weights and pricing match, but direct API infrastructure was not measured here.

Verdict

For standard production workloads, Gemini 4 Argon delivers results comparable to Opus 5.5, Astra, and Fable 5.1 at lower latency and lower cost. Through November 30, it is the cheapest flagship tested, and its tokenizer advantages keep it economical on multilingual text even under standard December rates. Its main weaknesses lie in unsubstantiated factual embellishments in structured legal drafting and softer prose output: in a separate blind Polish creative writing benchmark, it took 8th place with 81.8 points, trailing Claude Opus 5.5 at 90.1 points (detailed in the Gemini 4 Argon Polish language analysis). This evaluation reflects a single run conducted on October 13, 2026.

Frequently asked questions

Is Gemini 4 Argon better than Claude Opus 5.5?

Across these 23 structured tasks, performance was practically identical: 99.5% versus 99.4%. Argon was 3.4x cheaper and noticeably faster, while Opus 5.5 showed stronger nuances in non-English prose generation.

What is Gemini 4 Argon’s API pricing?

Through November 30, 2026, pricing is set at $2 per million input tokens and $10 per million output tokens. On December 1, pricing shifts to $4 input and $20 output.

How was Argon accessed for this benchmark?

The tests were conducted using OpenRouter (google/gemini-4-argon), which was operational from Poland on October 13, 2026.

Kinetyra